feat(sglang): map reasoning_effort to per-level thinking budgets

[pos-ei-don]

Description

The Go side already forwards the request's effort as Metadata["reasoning_effort"] (core/backend/options.go), but the sglang backend only reads enable_thinking. Every level therefore gets the same model-level thinking_budget, so low, medium and high produce identical reasoning — the level only switches thinking on. Separately, ApplyReasoningEffort only recognises minimal|low|medium|high; xhigh or max leave DisableReasoning unset, so with reasoning_default: off a request asking for more reasoning gets none (HTTP 200, no warning).

This PR:

• sglang backend: new model option effort_budgets:low=…;medium=…;high=…. _build_sampling_params uses the budget of the request's level; unlisted levels and requests without an effort keep thinking_budget. Parsing reuses _parse_thinking_budget (invalid entries are skipped with a warning, LoadModel never fails over it). A level listed in effort_budgets also turns reasoning on for that request. No behaviour change without the option.
• Go: ApplyReasoningEffort treats xhigh and max as thinking levels.
• Docs: effort_budgets in the SGLang reasoning section; xhigh/max in the reasoning_effort lists.

Same idea as vLLM's proposed --reasoning-effort-budgets (vllm-project/vllm#55432 / #55382) and Ollama's think levels (ollama/ollama#17566). Follow-up to #12193.

Notes for Reviewers

Measured on a DGX Spark (GB10), Qwen3.8-35B-A3B NVFP4, sglang 0.5.21, enable_strict_thinking, temperature 0, with effort_budgets:low=256;medium=1024;high=4096;xhigh=16384;max=16384:

Characters of reasoning per level:

prompt	before (every level)	low	medium	high	max
Pell equation x²−7y²=1, x<10000	629	629	2,036	2,882	2,882
3×10 domino tiling	—	—	4,013	10,211	37,981
divisibility proof (false as stated)	—	—	1,939	10,055	25,168

Before the change, low/medium/high were byte-identical and xhigh/max produced 0 characters of reasoning. On the false divisibility claim only max stated the correct condition; medium/high produced a wrong one.

Tests: four new unit tests in backend/python/sglang/test.py (parsing, per-level selection, fallback, map without default) plus one for the thinking toggle, run in the backend's own venv on the Spark: 29 tests, all new ones pass; test_build_prompt_forwards_enable_thinking fails identically with and without this change on that machine's sglang build. Reverting the selection line makes the new tests fail. The Go spec in core/config/reasoning_effort_test.go follows the existing cases but was not run locally.

Signed commits

• Yes, I signed my commits.
• Documentation updated (docs/content/) for user-facing changes, or not applicable

[localai-org-maint-bot]

@mudler Code-review sign-off at 0bd723a65285a710c1355fad8776f8c43c0b8b26: no blocking source findings; good to merge once CI validates it.

Reviewed against latest master 91c3bf084bd4952cc51ce6e114ba71dea604ade2. Per-level budget selection preserves the model fallback, explicit thinking-disable metadata takes precedence, and xhigh/max retain the operator-disable safeguard. The change includes documentation and focused regression cases.

Validation: git diff --check passes. I attempted go test ./core/config -run Test -ginkgo.focus='ModelConfig.ApplyReasoningEffort' -count=1 -timeout=30s, but compilation fails on stale generated protobuf bindings (pb.Animate3DRequest, pb.MotionRequest, and related symbols); the same failure occurs on unchanged master. The tests did not execute. Python is unavailable in this environment, so I could not run the SGLang suite. Only DCO is currently reported as passing; this sign-off does not establish runtime or CI readiness.

[pos-ei-don]

Thanks. For the record, since the bot could not run the suites: at 45a960c (adds two tests on top of the reviewed 0bd723a):

• Go: lint and tests-linux pass in a fork run of this branch.
• Python (backend/python/sglang/test.py) on a DGX Spark (GB10, sglang 0.5.21): 30/31 pass; the one failure, test_build_prompt_forwards_enable_thinking (SimpleNamespace lacks Images), fails identically on unchanged master.